DeepSeek Locked In a 75% Price Cut. Your Legacy Model IDs Still Expire in July.

Via DeepSeek: Models & Pricing

Cheap tokens are not a migration plan

DeepSeek has made its 75% price cut on V4 Pro permanent, according to The Next Web and the company's own API pricing page. Output that once cost up to $3.48 per million tokens now tops out at $0.87. Input on cache hits can run as low as $0.003625 per million tokens on V4 Pro.

For owner-led firms watching AI bills, that sounds like a win. It is — until you read the footnote. DeepSeek's docs state that legacy model names deepseek-chat and deepseek-reasoner will be deprecated on July 24, 2026 at 15:59 UTC — then fully retired and inaccessible. If your agents, scripts, or vendor integrations still call those IDs, you have a hard deadline whether or not you noticed the price drop.

The promotion was originally scheduled to end May 31, 2026. DeepSeek locked it in roughly a month after launching V4 Pro and V4 Flash, both with 1 million-token context windows. The company framed the move as the "era of cost-effective 1M context length." Competitors charge materially more: The Next Web cited OpenAI GPT-5 at $10 per million output tokens and Anthropic Claude Opus 4.7 at $25 output — against DeepSeek's $0.87 ceiling on V4 Pro.

Why the price war and the sunset land together

DeepSeek is buying market share with margin. That is rational vendor behavior. What is easy to miss is that price cuts and model retirements often arrive in the same quarter — you save money on tokens while your integration layer quietly rots.

Since the V4 preview shipped in April 2026, the legacy names have been aliases routing to V4 Flash (non-thinking and thinking modes respectively). Your app may already be on new weights without a code change — until July 24, when the old strings stop working entirely. Migration paths are documented: swap to deepseek-v4-pro or deepseek-v4-flash, keep the same base URL, and enable thinking mode via API parameters where needed.

V4 Flash is even cheaper — $0.14 input / $0.28 output per million tokens on cache miss, per DeepSeek's pricing table. V4 Pro sits at $0.435 / $0.87. Both support tool calls, JSON output, and up to 384K output tokens. The business question is not "which is cheaper?" but "which jobs need Pro reasoning versus Flash speed — and what happens when we swap?"

There is a geopolitical layer The Next Web and Reuters both note: routing sensitive workloads through a Chinese AI provider carries compliance and data-sovereignty questions that price alone cannot answer. Anthropic has publicly accused DeepSeek of "distillation attacks" — training on Claude outputs — an allegation DeepSeek has not fully addressed. None of that removes the July deadline. It complicates whether the cheapest capable model is the right default for your data.

What sensible firms do before July 24

• Inventory every model string. Search repos, env vars, LangChain configs, Zapier flows, and shadow scripts for deepseek-chat and deepseek-reasoner. The alias masked the change; the sunset will not.
• Test V4 Pro and V4 Flash in staging. Do not assume identical outputs. Re-run evals on summarisation, coding, and agent tool-calling before production cutover.
• Separate cost optimisation from continuity. A 75% discount is a finance win. A retired model ID is an outage. Track both in the same governance review.
• Document fallbacks. If DeepSeek is primary for low-cost long context, what is secondary for regulated or client-facing work? One vendor, one model string, one region is not a strategy.
• Re-price on the official page, not a blog post. Launch-window coverage listed higher V4 Pro figures; DeepSeek's live pricing page is canonical and has moved since April.

Token prices will keep falling across the industry in 2026. Model IDs will keep expiring. Treat both as operational events — not as background noise from the engineering team.

Someone has to own the model map — not just the invoice

DeepSeek's permanent discount makes model selection a CFO conversation and a continuity conversation at once. Cheaper inference only pays off if the swap does not break production on a Tuesday in July.

This is the work AgentsROI.ai does.

• Model Selection & Continuity Planning maps workloads to the right model tier, documents migration steps, and keeps fallbacks tested before vendors pull the rug.
• Managed AI Operations monitors which model IDs run in production, tracks vendor pricing changes, and catches alias retirements before they become outages.
• A Workflow ROI Audit shows where cheap long-context models save real money — and where privacy or quality requirements justify paying more elsewhere.

AgentsROI.ai is vendor-neutral. DeepSeek's whale is not your governance plan. The July 24 sunset is.

Lock in savings — and a fallback before the IDs go dark

DeepSeek just made the price war louder. It also put a date on the legacy model names your stack may still be calling. Smart owners celebrate the discount and schedule the migration in the same meeting.

Book a Model Selection & Continuity Planning session and find out which model strings your business actually depends on — before July 24 finds out for you.

This article summarizes publicly reported information from DeepSeek API documentation, The Next Web (May 2026), and related reporting. It is for general informational purposes only. It does not constitute legal, tax, financial, investment, security, or compliance advice. AgentsROI.ai is not a law firm, accounting firm, or registered investment adviser. Facts, pricing, statistics, and product capabilities cited here reflect the sources listed at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances—including applicable New York State and New York City requirements and any restrictions on use of foreign AI services. AgentsROI.ai may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product, model, or provider. References to DeepSeek or other companies describe reported pricing and product changes and do not constitute investment recommendations.